{
  "schema_version": "1.1",
  "id": "archive:https://arxiv.org/abs/2607.27146v1",
  "slug": "2607-27146v1-0gcx6ia",
  "url": "https://feed7.dev/p/2607-27146v1-0gcx6ia",
  "title": "MindForge: Teaching Small Language Models Whole-Life-Cycle Software Engineering via Source-Free Program Synthesis",
  "why_included": "MindForge trains a 27B coding model on source-free program synthesis, raising ProgramBench pass rate from 37.98% to 49.51% and improving across seven unseen benchmarks.",
  "summary": "MindForge turns open-source CLI programs into source-free training environments exposing only a compiled reference executable and documentation. Trajectories generated with **GLM-5.2** were used to fine-tune **Qwen3.6-27B**, lifting its ProgramBench average test pass rate from **37.98% to 49.51%**.",
  "practical_implication": "Builders should distinguish repository editing from whole-program creation when evaluating coding agents. Training against executable behavior and documentation may offer a scalable route to agents that can generate repositories, translate code, fix bugs, and implement features without seeing reference source.",
  "agent_context": "MindForge turns open-source CLI programs into source-free training environments exposing only a compiled reference executable and documentation. Trajectories generated with **GLM-5.2** were used to fine-tune **Qwen3.6-27B**, lifting its ProgramBench average test pass rate from **37.98% to 49.51%**.\n\nBuilders should distinguish repository editing from whole-program creation when evaluating coding agents. Training against executable behavior and documentation may offer a scalable route to agents that can generate repositories, translate code, fix bugs, and implement features without seeing reference source.\n\nThe results come from the authors’ benchmark suite rather than production deployments. Frontier models still fully resolve **fewer than 1%** of ProgramBench tasks, and the material does not establish how the pipeline handles graphical apps, services, or undocumented behavior.",
  "source": {
    "name": "arXiv",
    "url": "https://arxiv.org/abs/2607.27146v1",
    "published_at": "2026-07-29T17:23:02.000Z"
  },
  "source_class": "blog_post",
  "content_type": "Paper",
  "layer": "agent",
  "domains": [
    "coding"
  ],
  "topics": [
    "coding-agents",
    "harness-engineering",
    "agent-evals"
  ],
  "verification": {
    "status": "needs_review",
    "label": "Needs Review",
    "method": "unverified",
    "verified_at": null
  },
  "uncertainty": [
    "The results come from the authors’ benchmark suite rather than production deployments. Frontier models still fully resolve **fewer than 1%** of ProgramBench tasks, and the material does not establish how the pipeline handles graphical apps, services, or undocumented behavior."
  ],
  "connected_context": {
    "meaning": "This expands coding-agent training beyond editing visible repositories to reconstructing whole programs from documentation and executable behavior. It confirms that environment and trajectory design can materially improve a smaller model, while the sub-1% full-resolution rate sharply limits claims of autonomous software creation. Production relevance remains open because the benchmark excludes important application classes and undocumented behavior.",
    "corpus_size": 263,
    "generated_at": "2026-07-30T10:53:29.906Z",
    "connections": [
      {
        "title": "The Prime Intellect Stack — Will Brown, Prime Intellect",
        "source_name": "YouTube",
        "source_url": "https://www.youtube.com/watch?v=V-EDrhIhHzQ",
        "feed7_url": "https://feed7.dev/p/the-prime-intellect-stack-will-brown-prime-intellect-1dc0rrp",
        "reason": "MindForge is a concrete instance of the environment-centered eval, data-generation, and training stack described by Prime Intellect, using executable programs as composable behavioral targets."
      },
      {
        "title": "Harness design for long-running application development",
        "source_name": "Anthropic",
        "source_url": "https://www.anthropic.com/engineering/harness-design-long-running-apps",
        "feed7_url": "https://feed7.dev/p/harness-design-long-running-apps-15vc0wu",
        "reason": "Both use external behavioral evaluation to counter unreliable self-assessment, but Anthropic applies it during long-running app construction while MindForge uses it to generate training trajectories and benchmark whole programs."
      },
      {
        "title": "Effective harnesses for long-running agents",
        "source_name": "Anthropic",
        "source_url": "https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents",
        "feed7_url": "https://feed7.dev/p/effective-harnesses-for-long-running-agents-0xzfs05",
        "reason": "The multi-session harness supplies incremental state and verification for repository work; MindForge addresses a different prerequisite by training models to create whole programs from documentation and observable behavior."
      },
      {
        "title": "Harness Engineering is not Enough: Why Software Factories Fail — Dex Horthy, HumanLayer",
        "source_name": "AI Engineer",
        "source_url": "https://www.youtube.com/watch?v=Ib5GBkD555M",
        "feed7_url": "https://feed7.dev/p/harness-engineering-is-not-enough-why-software-factories-fail-dex-horthy-1tj1amt",
        "reason": "MindForge’s pass-rate improvement measures functional output, while the software-factory critique identifies maintainability and affordable human review as additional requirements not established by benchmark success."
      }
    ]
  },
  "lifecycle": "Current",
  "published_at": "2026-07-29T17:23:02.000Z",
  "modified_at": "2026-07-29T17:23:02.000Z",
  "supersedes": [],
  "expires_at": null,
  "formats": {
    "html": "https://feed7.dev/p/2607-27146v1-0gcx6ia",
    "json": "https://feed7.dev/p/2607-27146v1-0gcx6ia.json",
    "markdown": "https://feed7.dev/p/2607-27146v1-0gcx6ia.md"
  }
}